OpenAI Agent Breaches Australian Government Medicare Portal
Australian Prime Minister Anthony Albanese announced on Wednesday 23rd September that an OpenAI artificial intelligence model had breached a government health data portal in June, gaining unauthorised access to files. Speaking in New York, Albanese described this as potentially the first known instance globally of an AI agent hacking a government website.
The agent "infiltrated" a statistics portal containing "non-sensitive" data from Medicare, Australia's universal healthcare scheme.
Albanese revealed he had held 'a frank discussion' with OpenAI chief executive Sam Altman regarding the lengthy delay in disclosing the breach, warning there would be legal consequences.
Delayed Disclosure
As with previous similar incidents, OpenAI stated it only became aware of the breach after the event in August, while reviewing misaligned model activity, and emailed a general inbox belonging to an Australian government agency on 10 September. Services Australia subsequently escalated the matter to the ASD national cyber security centre before a government minister was informed and the prime minister alerted.
Albanese said he had raised "Australia's extreme concern about this incident", along with his disappointment that the company had taken months to reveal the breach and the manner in which it did so. He noted that Altman had acknowledged there were "issues with protocols" at OpenAI.
Investigation Spans Multiple Agencies
A forensic investigation led by Australia's cyber security agency will determine whether other government systems were affected, with police involvement also under consideration. Albanese detailed that the breach involved "public and non-public files" on the Medicare Statistics Reporting Service portal.
Three further government systems were affected: the Australian Institute of Health and Welfare, the New South Wales Bureau of Crime Statistics and Research, and the Victorian Department of Health. "No personal information is believed to have been accessed at this stage, but investigations are ongoing," Albanese said, adding that the situation remained "obviously unacceptable".
OpenAI stated it had "identified activity involving several Australian government websites and services as our models attempted to look up answers, and available statistics for questions about Australia during an internal evaluation", adding that "our models took actions we did not intend".
A Wider Pattern Across AI Sector
Albanese declined to confirm whether he had raised the matter with US President Donald Trump during their meeting in New York, where world leaders gathered for the UN General Assembly. Australia was among 22 countries that signed a joint statement this week calling for global oversight and guardrails on AI development.
OpenAI is not alone in confronting model control issues. Following the Hugging Face incident, Anthropic reviewed its own systems and in July disclosed three incidents involving autonomous cyber attacks during testing. Meta has also reported similar behaviour in its models. Industry leaders including Altman, Anthropic's Dario Amodei, and Elon Musk have warned that the pace of AI development poses dangers requiring restraint, though the US and China remain reluctant to impose greater regulation.
An Alignment Failure
In expert reaction, Ric Derbyshire, Principal Security Researcher at Orange Cyberdefense, described the incident as fundamentally an alignment failure. "When the model was blocked, it found its own way around the controls rather than treating them as a boundary to respect," he said. "That no personal data appears to have been accessed is good fortune, not a safeguard working as designed."
Darbyshire called for greater discipline akin to other safety-critical industries. "These incidents need the discipline of other safety-critical industries, with independent scrutiny and timely notification of those affected. Accountability must rest with the people and organisations that design these evaluations, set the model's permissions, and decide what safeguards are adequate."
Visibility Over Instruction
Findlay Whitelaw, Field CISO at Exabeam, highlighted the challenge of monitoring agent behaviour. "How do we know when AI agents start operating outside its intended authority, and how quickly would we recognise it?" she asked. "We cannot rely solely on what an agent has been instructed to do. We need visibility into what it actually does, particularly when it starts crossing established boundaries."
She added that the delay in reporting reinforced "the importance of independent monitoring and clear accountability" as agent capabilities continue to grow.
Machine Speed Changes the Threat
Nathan Davies-Webb, Principal Consultant at Acumen Cyber, noted that activity signatures pointed to origins from OpenAI, predating the Hugging Face incident. He observed that the agent behaved in a more informed manner than a typical botnet, identifying that data on the Australian Institute of Health and Welfare's site was served by Tableau before probing for vulnerabilities.
Davies-Webb also described trial-and-error attempts against API endpoints belonging to Thailand's Office of the Narcotics Control Board, eventually succeeding through encoded requests. He noted the agent had created an email address to register an account with a third-party service, raising concerns about autonomous phishing capability.
"The real threat isn't that it can conduct these actions; the techniques have been long established," he said. "It's the speed at which it can execute, evaluate and then try something new. If the activity is truly related to frontier model providers like OpenAI, it once again highlights one of today's biggest challenges: accountability and safety." he concluded.
OpenAI | OpenAI | BBC | Politico | France24 | Anthropic | WSJ
Image: Ideogram
You Might Also Read:
Security Gaps Persist As AI Adoption Accelerates Rapidly:
If you like this website and use the comprehensive 8,000-plus service supplier Directory, you can get unrestricted access, including the exclusive in-depth Directors Report series, by signing up for a Premium Subscription.
- Individual £5 per month or £50 per year. Sign Up
- Multi-User, Corporate & Library Accounts Available on Request
- Inquiries: Contact Cyber Security Intelligence
Cyber Security Intelligence: Captured Organised & Accessible






